Advanced analytics

Four advanced statistical analyses, computed automatically from open data, with settings fixed in advance and their limits shown.

43.8 % share of "Very hot days in the year" in contributions

Updated : 23 October 2026

What does the model use to reproduce the financial stress index?

A random forest (500 trees) reproduces the financial stress index for each year (1999-2025) from five variables. Shapley values split, year by year, the gap between the model's value and the average level (61.1) among these variables.

Overall importance (mean absolute contribution)

Very hot days in the year 43.8 % · ±6.5
Previous year's FSI 33.2 % · ±4.9
Agricultural value added growth 10.9 % · ±1.6
Annual temperature anomaly 7.3 % · ±1.1
Annual drought (December SPEI-12) 4.8 % · ±0.7

In-sample R²0.58

Out-of-sample error (RMSE, 17 years)22.4

Naive model error (previous year's value)13.2

Out of sample, the model forecasts worse than simply repeating last year's value: the contributions below describe how it reproduces the past, not a forecasting tool.

Contributions year by year

Choose a year

Average level61.1

Very hot days in the year (65.4 d) +4.1

Previous year's FSI (50.5) −7.7

Agricultural value added growth (3.7 %) +0.8

Annual temperature anomaly (1.1 °C) −0.8

Annual drought (December SPEI-12) (0.4) −0.9

Model value56.7

Observed value50.0

How to read: a positive contribution means the variable's value that year led the model to place the index above its average level. These contributions describe how the model works and establish no causal effect. The sample is short (about thirty years), and a variable that rises over time, such as very hot days, may capture a common trend rather than a direct relationship.

Other analyses on the site

Unusual climate months Isolation Forest

6 cases in September 2026, including Laayoune-Sakia El Hamra: many very hot days.

Details on the Alerts page → Financial stress index forecast Random forest · Gradient boosting

Out-of-sample errors (RMSE): forest 16.2, boosting 19.4, naive model 13.2.

Compare models → News sorted by theme Multilingual embeddings

Each release is automatically sorted by theme, with a short summary extracted from its text.

News page →

Method and limits

All computations run during the daily update with scikit-learn, without any external service, with fixed hyperparameters and random seed: results are reproducible.
Shapley values are computed exactly (all coalitions of the five variables), without approximation.
Tree models learn from the past and do not extend trends: they cannot forecast a value outside the range already observed.
Prediction is not causation: a variable weighing in the model does not mean it causes the index to change.

Source : Calculs de l'Observatoire (scikit-learn) ; Open-Meteo (ERA5) ; Banque mondiale